Re: false positives from git diff-index when used with --git-dir

From: Alexander Gladysh <agladysh@gmail.com>
Date: Oct 14, 2010, 17:20 UTC
Message-ID: <AANLkTimqvmecpt_j=HqHF1tircO5bPJHtRW-ePQ-wE+T@mail.gmail.com>
In-Reply-To: <20101014170227.GA14429@sigill.intra.peff.net>

On Thu, Oct 14, 2010 at 21:02, Jeff King <peff@peff.net> wrote:

> On Thu, Oct 14, 2010 at 06:40:51PM +0400, Alexander Gladysh wrote:

>> I'm *sometimes* (once in about ten times I use it -- repository
>> changes between invocations of course) getting false positives from
>> git diff-index when I invoke git from another directory:

>> git --git-dir=path/to/.git --work-tree=path/to update-index -q --refresh
>> git --git-dir=path/to/.git --work-tree=path/to diff-index --exit-code
>> --quiet HEAD

>> This is *really* annoying and really kills the benifits from my
>> workflow automation scripts.

>> Unfortunately I can't find a reproducible use case for this bug.

>> Is there anything I can do to help catch it?

> Can it reproduce over a large number of trials? I.e., something like:

>  while true; do
>    git ... update-index ...
>    git ... diff-index ... || echo failed
>  done

> If so, try adding a "sleep 1" between the two commands. If that fixes
> it, it implies a race condition in git.

No, I'm not able to reproduce it this way. (Perhaps some more complex operations on git tree are required -- looks like it happens only if tree was changed.)

> You could also try running under valgrind to see if there are any bad
> memory accesses, which could also produce intermittent buggy behavior.

Thanks, I'll try it.

It feels like a race condition somewhere. (Quite possibly even in my automation tool -- since I'm not able to reproduce bug outside of it -- if so, sorry for the noise.)

I'll try some more debugging and will get back to you if I will have any new information.

Thank you, Alexander.